Comparing confidence-guided and adaptive dynamic pruning techniques for speech recognition

Improvement in pruning algorithms for automatic speech recognition leads directly to a more efficient recognition process. Efficiency is a very important issue in particular for embedded speech recognizers with limited memory capacity and CPU power. In this paper we compare two pruning algorithms, the confidence-guided pruning and the adaptive control pruning technique. Both methods set the pruning threshold for the Viterbi beam search process dynamically for each time frame depending on search space properties. We show that both dynamic pruning techniques are applicable in reducing the time consumption of the recognizer whereas our novel confidence-guided pruning approach outperforms the adaptive control technique clearly.